Last week I attended the annual OCP Summit in San Jose, where all the cool hardware kids hang out. Surprisingly often, when I am explaining that DriveScale makes server infrastructure more flexible and efficient, I am asked whether we are based on OCP technologies. Short answer: No. Not at all. We exist in different planes of reality.
OCP is a rather odd forum. It was started by Facebook to “open up” Facebook’s hardware designs and requirements. Of course, every potential vendor to Facebook wanted to participate, and then Microsoft joined as another major sponsor – so now lots of hardware issues are discussed and some interesting standards have emerged. The summit has a show floor where hardware vendors show off their latest & greatest offerings, and where Facebook and Microsoft continue to show off what they’re using and what they want.
Approximately half of all IT equipment is now being sold to hyperscalers like Amazon, Microsoft, Google, Facebook, etc. Any *component* (CPU, SSD, etc) level manufacturer must capture business from them to be viable. But most of the OCP activities revolve around *system* level issues – how wide is your rack, how dense, how complex is the cabling, etc. The output of OCP for these issues are good for the hyperscalers but are not things that can be easily adopted in most other organizations.
It was Jason Taylor of Facebook, way back in 2013, who first preached the gospel of “disaggregation” – allowing simpler SKUs and independent life cycle management for data center hardware. So now OCP is full of disaggregated technologies – disk, SSD, even GPUs. But the hyperscalers don’t really do dynamic composability yet – they still rely on manual cabling to tie together the disaggregated components. They don’t need the flexibility of composable infrastructure because they have huge teams understanding their workloads, doing capacity planning, and managing the eager vendors in Taiwan who can turn around product requests in 2 months.
But you, dear reader, are not Google or Amazon or Facebook. Even if you succeed in understanding your workload, predicting where it is going is a black art. Nor do you get to tell your vendors that you want custom features for no added cost. Composable Infrastructure is for you – to allow you to react, re-provision, and re-optimize as your demands change. You even get to create “what if” clusters without breaking the bank.
All of the amazing components shown at OCP will be available through the usual vendors to normal scale datacenters. Composability means you can attach cool new stuff like NVMe drives to your older servers – you’re not limited by sheet metal any more. And component vendors are now becoming supporters of composability. At OCP we saw Western Digital (already a strong DriveScale partner) showing off their up and coming OpenFlex Composable Infrastructure product (we already have one in the lab); Marvell was touting composability with their Ethernet switch chips and NVME-oF target chips; and Mellanox showed their new BlueField SNAP product which is also targeted for Composable Infrastructure. We’re eager to work with all these guys, and more. DriveScale has great influence with hardware companies because of my co-founder Satya Nishtala, who has been a leader and mentor in the Silicon Valley hardware industry for 30+ years – he has friends and disciples everywhere!
So back to the big question: OCP – Will It Compose? Sure it will! But you don’t need it for Composability. OCP and Composable Infrastructure are independent approaches to data center optimization.